Skip to content

[ECCV 2024] Sparse Beats Dense: Rethinking Supervision in Radar-Camera Depth Completion

License

Notifications You must be signed in to change notification settings

megvii-research/Sparse-Beats-Dense

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

[ECCV 2024] Sparse Beats Dense: Rethinking Supervision in Radar-Camera Depth Completion

Huadong Li*, Minhao Jing*, Jing Wang, Shichao Dong, Jiajun Liang, Haoqiang Fan, Renhe Ji

MEGVII Technology

*Equal contribution Lead this project Corresponding author



News

  • [🔥🔥🔥 2024.07.18] Code is released.

Overview

Model

It is widely believed that sparse supervision is worse than dense supervision in the field of depth completion, but the underlying reasons for this are rarely discussed. To this end, we revisit the task of radar-camera depth completion and present a new method with sparse LiDAR supervision to outperform previous dense LiDAR supervision methods in both accuracy and speed.

Specifically, when trained by sparse LiDAR supervision, depth completion models usually output depth maps containing significant stripe-like artifacts. We find that such a phenomenon is caused by the implicitly learned positional distribution pattern from sparse LiDAR supervision, termed as LiDAR Distribution Leakage (LDL) in this paper. Based on such understanding, we present a novel Disruption-Compensation radar-camera depth completion framework to address this issue. The Disruption part aims to deliberately disrupt the learning of LiDAR distribution from sparse supervision, while the Compensation part aims to leverage 3D spatial and 2D semantic information to compensate for the information loss of previous disruptions.

Preparation

  • Environment

    • Detailed environment settings should be found with requirements.txt
pip3 install -r requirements.txt
-  data
  - nuscenes_radar_5sweeps_infos_test.pkl
  - nuscenes_radar_5sweeps_infos_train.pkl
  - nuscenes_radar_5sweeps_infos_val.pkl
  - nuscenes
     - samples
     - seg_mask.tar

cd data/nuceneses
tar xvf seg_mask.tar
  • Pretrained weights

mkdir checkpoints & cd checkpoints
mv model.ckpt ./chekpoints/model.ckpt

Training

OMP_NUM_THREADS=4 torchrun --nproc_per_node ${GPU Nums} train_DDP.py

Eval

python3 eval.py -m ./checkpoints/model.ckpt

BibTeX

@misc{li2023sparsebeatsdenserethinking,
      title={Sparse Beats Dense: Rethinking Supervision in Radar-Camera Depth Completion}, 
      author={Huadong Li and Minhao Jing and Jiajun Liang and Haoqiang Fan and Renhe Ji},
      year={2023},
      eprint={2312.00844},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2312.00844}, 
}

Contact

If you have any questions, feel free to open an issue or contact us at lihuadong@megvii.com or jirenhe@megvii.com.

About

[ECCV 2024] Sparse Beats Dense: Rethinking Supervision in Radar-Camera Depth Completion

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages